Silicon Investor (SI) -- The First Internet Community

STOCKTALK

We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor. We ask that you disable ad blocking while on Silicon Investor in the best interests of our community. If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.

Strategies & Market Trends : Neural Nets - A tool for the 90's -- Ignore unavailable to you. Want to Upgrade?

To: LastShadow who wrote (738)	4/5/2000 10:29:00 AM
From: Optim	Read Replies (3) \| Respond to of 871

Lastshadow, You've hit the nail on the head with this post. You've brought up most of the issue I've grappled with, and I'm excited to participate in the exercise. But, on a general level, that could be a good excercise for the thread, to develop a testing process, a methodology that structures the development of a net and the steps one would take to refine it. That would allow the lurkers and posters here to contribute and learn also. I agree. I am willing to post all efforts on my part to the thread, and possibly a website if charts/downloads are needed. Lets pick a sector - my vote is for Tech due to the volatility, trading ranges and volume Tech is fine with me. Do we want to whittle down to a specific industry as well? If so, I had good success with semiconductors, biotechs and oil/gas, although oil/gas isn't a tech. Like you said, subliminally we all influence our models... :) Then lets pick an objective: minimize error, maximize profit, accuracy of prediction - I would go for any of them, but if we go with Tech sector, we should probalby forget minimizing error, as the trading strategy we couple with this net would be sector-specific, and the erratic nature means one would use stops rather than a net to handle that concern. I'm all for Maximize Profit. Im my tests, minimizing error and accuracy of predicition don't work as well unless they are combined with a good trading strategy on top. Too many small moves will give you good accuracy or low error, but make the net unprofitable. Maximising profit has seemed to work best for me, although trades can often be too long for my liking. This can be compensated for in the trading strategy. Should we pick a single stock, a few stocks or an index or spyder? I vote for a set of stocks, just because I would want to determine fitness and eliminate spurious data by looking at the individual tickers rather than some masked reaction in an index or spyder, but thats just my vote This brings up another issue. In my testing, shorter periods of training data yeilds better profits, in sample and out of sample. I often worry about about overfitting the training data though. By using multiple stocks that have been screened to be relatively similar we can increase the amount of data out nets are exposed to, without increasing our training periods. Something that holds up across stocks will likely continue to hold up out of sample, and in realtime. What I would like to hear from others is what they would like to get out of this - the object in my mind would be to help learn How to develop nets. Appreciate that this is not a quick and straightforward process, as there are a lot of qualifications and assessments made along the way that would be revisited each time wone looks at a new net I'd agree in that it can be viewed as a set of guidelines for the construction of a neural network in combination with some form of a trading strategy. The two often complement each other. I'd like to be able to build a document to outline the steps involved in every step of the process. This would include the collection of relevant data, analyzing the data to find suitable transforms/preprocessing, training and construction of the networks and finally testing the output. I'm sure the process will be repeated as we analyse our results and go back to modify the various steps. One thing I think we should agree upon is the platforms used. I have the following tools available, along with two good computers (dual processors) to use them on: Neuroshell DayTrader, BioComp Profit, Metastock, Excel, WaveWi$e, and if need Genehunter a GA package. I can also get tick data if it helps to decide if trading on the open or close is better. I know Lastshadow has access to NS Trader, so perhaps it might be a suitable package? One other issue might be timeframe. I'm all for end of day position trading, as it might be easier to model. So then, how shall we commence? Optim

To: LastShadow who wrote (738)	4/5/2000 6:12:00 PM
From: xu, b.	Respond to of 871

But, on a general level, that could be a good excercise for the thread, to develop a testing process, a methodology that structures the development of a net and the steps one would take to refine it. That would allow the lurkers and posters here to contribute and learn also. Bai: good idea! So, if some others are willing, we can start this process with a scope: Bai: again good idea! Lets pick a sector - my vote is for Tech due to the volatility, trading ranges and volume Bai: agree. Then lets pick an objective: minimize error, maximize profit, accuracy of prediction - I would go for any of them, but if we go with Tech sector, we should probalby forget minimizing error, as the trading strategy we couple with this net would be sector-specific, and the erratic nature means one would use stops rather than a net to handle that concern. Bai: I would pick P/R ratio as the output. P/R stands for profit/risk. Should we pick a single stock, a few stocks or an index or spyder? I vote for a set of stocks, just because I would want to determine fitness and eliminate spurious data by looking at the individual tickers rather than some masked reaction in an index or spyder, but thats just my vote. Bai: agree. What I would like to hear from others is what they would like to get out of this - the object in my mind would be to help learn How to develop nets. Appreciate that this is not a quick and straightforward process, as there are a lot of qualifications and assessments made along the way that would be revisited each time wone looks at a new net. Bai agree.